Decentralized Multi‐Graph Based on Attention Network for Social Recommendation

ABSTRACT The user's interest preference in a recommender system will be influenced by their own historical behavior and social relationships at the same time, recommendations based on social relationships between users, and between users and items have become the dominant trend. However, effectively modeling these heterogeneous relational signals remains a challenging problem. Many existing graph‐based recommendation approaches rely on a unified graph structure to represent multiple types of relationships, which may introduce information interference and limit the ability to capture diverse structural dependencies. In addition, rating data often contain inherent biases caused by user rating habits and item popularity, which may negatively affect preference modeling. In order to more effectively use user social relations for social recommendation, the graph neural network is the key method for its implementation. This study proposes an attention network‐based decentralized multi‐graph social recommendation algorithm (DMSR) based on the traditional graph neural network model, which integrates the bias of users' rating information and homogeneous items. The model integrates the bias of user ratings and reduces the preference error through the decentralized multi‐graph structure. At the same time, the algorithm introduces a homogeneous item aggregation module to strengthen the perception of items, and utilizes an attentional mechanism to comprehensively consider the influences from different friends and similar items. Finally, the model has been experimentally validated based on two real datasets, which confirms that DMSR has better recommendation performance than the baseline social recommendation algorithm with high robustness.